The runswift infrastructure is fairly large, so this page lists some key aspects and summaries. The best way to understand these in detail is to read the reference papers and relevant code.

Overview - libraries, libagent, runswift and offnao

libagent is a library loaded into naoqi and how runswift communicates with the robot hardware on Nao v5 and earlier (OS 2.1 and earlier).

soccer & soccer-static are the basic toolchain-independent modules (perception without cameras, etc.). The static version is for robot-static, and the dynamic version is for [[offnao]].

robot-static is soccer-static with hardware-specific modules (camera, some parts of motion).

runswift is robot-static with main().

[[offnao]] is our current C++/Qt visual debugger that connects to runswift via TCP port 10125. It can also save and load recordings in a few different formats

If you poke around the cmake files, they should correlate with what’s said here, but if it doesn’t, the cmake files take precedence (and please update this).

Drivers and libagent

Not sure about drivers? Naoqi modules like libagent are loaded in image/etc/naoqi/autoload.ini

If you poke around the files, they should correlate with what’s said here, but if it doesn’t, the files take precedence (and please update this).

Nao v5, OS 2.1, and earlier

libagent is our communications bridge to the robot. It is a library that we wrote within Aldebaran’s proprietary NaoQi SDK that receives information such as joint angles, and sends information such as LED commands. It is also responsible for button presses. libagent communicates with our executable (runswift) via a shared memory object. (linux allows you to allocate named memory that is shared with other processes.)

Nao v6, OS 2.8, and later

LoLA is Aldebaran’s communications bridge to Aldebaran’s proprietary NaoQi SDK. lola is a process separate to naoqi. Access is only available on special RoboCup versions of the OS. To enable access, you need to have /home/nao/robocup.conf on the Nao. This file can be empty. lola will then create a socket file /tmp/robocup which supports msgpack for read and write. See robot/motion/LoLAData.cpp for more details.



There are many options to the runswift executable. They can be put on the command line, in the configuration file runswift.cfg, in the per-robot configuration file , or sent from off-nao. They are configured in options.cpp, and viewable by running “runswift –help”.

Call tree

At the top of our call tree is main, naturally. main.cpp launches threads for most of our top level-modules, such as Perception and Motion. The threads continuously run in a loop, independent of each other. Perception is [currently] the only top-level module that contains other top-level modules, namely Kinematics, Vision, Localisation, and Behaviour, which are executed in that order. Behaviour contains a bridge to python, and calls the execute method of a top-level python skill ( skill by default).


The threads are run in a function called safelyRun, which:

  • restarts the thread in the event of a segfault or other fatal signal

  • restarts the thread in the event of an exception being thrown

  • monitors for the thread if it goes overtime

  • sleeps the thread if it goes undertime (to allow other threads to run)

  • [re]constructs the thread, repeatedly calls the thread’s tick() function, and destructs the thread safelyRun stop calling tick() when ctrl-c is pressed or the interrupt signal is received. Additionally, perception is monitored for freezes (tick() does not return).

To decrease the need for locks, and decrease the possibilities for segfaults, most of the blackboard variables are not pointers, and exist as long as the blackboard exists. It would be a concurrency error to have one thread write a new pointer to the blackboard and free the old one, while another thread may be accessing its contents. Some blackboard variables are smart pointers to avoid this.

Multiple threads should never write to the same blackboard variable. If two threads ever happened to write simultaneously, it could crash either/both threads, as well as leaving corrupt data, causing another thread to crash. Each sub-blackboard is written to only by its corresponding top-level module.

Note that reads of structures on the blackboard greater than 1-word in size, are not necessarily atomic, as the read may be interleaved with another thread’s write. An example is when reading from blackboard->


The blackboard is a global variable where threads share information. The blackboard is broken into sub-blackboards for each top-level module. Each sub-blackboard is written to by its corresponding top-level module. Variables on the blackboard should not contain pointers. Adding a variable to the blackboard does not make it immediately accessible in offnao. You have to manually archive it, see [[Blackboard Serialization]].


We have a cout-like interface for recording log messages to a file. For example:

llog(INFO) << "RUNSWIFT soccer library spinning up!" << endl;

Log files are saved in /var/volatile/runswift on the robot, and there is one log file per source directory. The log levels available are SILENT, QUIET, FATAL, ERROR, WARNING, INFO, VERBOSE, DEBUG1, DEBUG2, and DEBUG3. Setting a log level will record messages of that level or higher. The default log level is SILENT, and can be overridden using one of the options mechanisms listed in the options section. e.g. “-l INFO” will record messages at level SILENT, QUIET, FATAL, ERROR, WARNING, or INFO. Using llog (especially in tight loops) takes resources, even if the set log level is higher (e.g. “-l SILENT”) than the one passed to llog (e.g. “llog(DEBUG3)”).


Including the blackboard in every file that needs to access a blackboard variable has led to some unreasonably long compiles in previous years. To alleviate this, we are using an Adapter-like pattern, where only the files representing the top-level modules access the variables on the blackboard. Blackboard.hpp should only be included from these special <module>Adapter.cpp files, and the blackboard should only friend these Adapters.

Python Infrastructure

We use python as a high level language for writing behaviour code. The reasons for this are discussed in depth at: Carl Chatfield Robocup Report 2012 - Chapter 3

In order to enable this, we use the boost python libraries to set up an embedded python interpreter within our runswift executable that has access to parts of our blackboard. In every perception cycle, an entry tick() function is called by our c++ code within the python interpreter instance with a reference to the blackboard. The expected return value is a BehaviourRequest object that contains any information to pass back into the C++ code (primarily actions for motors & leds).

Key directories relating to the python interface:

  • robot/perception/behaviour/

    This folder contains the c++ parts of the behaviour chain. In BehaviourAdapter.cpp, it creates a PythonSkill (defined in the python subdirectory). On every tick, it executes the PythonSkill which returns a BehaviourRequest to it. This is then used to write actioncommands to the blackboard.

    • python/

      The important pieces here:

      • wrappers - Wrappers over some data types (see the report/code for details).

      • converters - Converters for arrays, etc (see the report/code for details).

      • PythonSkill.cpp - The PythonSkill class manages executing a python interpreter (setting up modules, paths, etc). It also watches files and reloads the interpreter if they change. This is useful for quick iteration on code (simply nao_sync your new code over and the robot runs the new code). It also handles python exceptions that are uncaught by flashing the Leds, saying “Python error”, and reloading the code. This is useful because in game, if your python code reaches an untested state and crashes, you want the interpreter to restart and continue (but notify you that it failed).

      • RobotModule.cpp - This pulls all the wrappers in and gets compiled into a python module which can be imported within the interpreter as robot. You can then access parts of the wrapped cpp code. e.g

# This is a simple
import robot

def tick(blackboard):
    req = robot.BehaviourRequest() # Creates a Behaviour Request instance.
    req.actions.leds.rightEye = robot.rgb(True, False, False) # Set right eye to red.
    return req
  • image/home/nao/data/behaviours

    This is where the python files that run on the robot are kept. This folder gets synced to the robot by nao_sync. The highest level file here is This file is run by PythonSkill at the top level. It calls the tick() function on this module, passing it a reference to the blackboard. The function must return a BehaviourRequest instance. Whatever else happens is up to your python architecture and can be customised to match some form of state machine / decision tree / other behaviour system implemented in python.